A $2.55 API Receipt Beats a Cheap Token Headline

01 / Price the whole request, not the cheapest token

Before your first API relay top-up, build a receipt. A low input-token price tells you little if your workload generates long answers, repeatedly sends large contexts, or pays separately for cache writes. The useful question is not whether a rate looks cheap. It is whether you can explain the final charge.

Start with four categories: ordinary input, output, cache reads, and cache writes. Check whether the published unit is per thousand or per million tokens. Then check whether cached tokens replace ordinary input charges or appear alongside them. Counting the same tokens twice can ruin your estimate before you send a request.

02 / A hypothetical workload totals $2.35 before retries

The following rates are invented solely to demonstrate the arithmetic. They are not CheapRelay prices, a live quote, a benchmark, or a promised discount. Each quantity represents a separate, non-overlapping billing category.

• ■ Ordinary input: 2 million tokens at $0.50 per million = $1.00.
• ■ Output: 400,000 tokens at $2.00 per million = $0.80.
• ■ Cache reads: 5 million tokens at $0.05 per million = $0.25.
• ■ Cache writes: 500,000 tokens at $0.60 per million = $0.30.
• ■ Workload subtotal: $1.00 + $0.80 + $0.25 + $0.30 = $2.35.

Ignoring caching would hide $0.55, or about 23% of this subtotal. That does not make caching bad value: assessing its benefit requires comparing the same workload without caching. It does mean that a receipt with only input and output lines is incomplete for this example.

03 / Two retry lines bring the receipt to $2.55

Now suppose additional attempts consume 200,000 ordinary input tokens and 50,000 output tokens at those same illustrative rates. Input adds $0.10; output adds another $0.10. The final hypothetical charge becomes $2.55.

This is not an assumed retry rate. It is a separate usage allowance showing why application behavior belongs in your budget. Ask how failed, interrupted, and partially completed requests are billed. Check your client's retry settings, too. An automatic second attempt can create another charge if it performs billable work.

Keep a small test log containing the request identifier, selected model, reported token categories, and balance change. Reconcile several requests individually before trusting an aggregate dashboard. Record displayed precision so rounding does not look like unexplained spending.

04 / A $5 top-up needs balance rules, not optimism

If a relay credited exactly $5 and deducted exactly our $2.55 example, $2.45 would remain. That arithmetic assumes no payment fee, currency conversion, tax, expiry, or other adjustment. It does not establish any provider's actual terms.

• ■ Confirm the minimum deposit and the amount actually credited.
• ■ Check payment fees, billing currency, and applicable taxes.
• ■ Read balance-expiry and unused-credit refund rules.
• ■ Check whether spending caps and low-balance alerts exist.
• ■ Confirm whether your required model has a separate rate schedule.

A small initial deposit limits the amount committed while you test. It does not replace checking the terms, and it is not necessarily refundable.

05 / Test CheapRelay's routes before funding real traffic

CheapRelay offers pay-as-you-go access through one key, with no monthly fee or subscription. Its billing covers real token usage, including cache reads and writes. Use the current rate for each applicable category rather than substituting the illustrative prices above.

The API base is https://api.cheapapirelay.com. OpenAI-compatible routes sit under /v1, while Gemini-native routes sit under /v1beta. CheapRelay works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients. Still, test the features your application actually uses, including streaming, tool calls, error handling, and usage reporting.

Sign up for CheapRelay, check the current rates and balance terms, and make your first top-up a controlled billing test. Move real traffic only after you can reconcile the receipt—not merely because the headline price looks attractive.